notes
book review: the creative process
24 may 2004
While the fifty-year history of Artificial Intelligence has witnessed much work on representing and reasoning with the foundations of cognition such as the modeling of beliefs and goals, relatively little work has been focused on higher-level cognitive phenomena such as socialisation, creativity and writing. Researching in the script-based story understanding traditions of Roger Schank and Mike Dyer, Scott Turner, in The Creative Process, presents a computational model of creativity and storytelling. The problem space is very interesting and fruitful, and there is still a paucity of research at these cognitive granularities.
Turner motivates the computational modeling of creativity with the goal of producing short 200-word-long computer-generated stories within the King Arthur domain. This program, called MINSTREL, was the topic of Turner's dissertation work with Mike Dyer at UCLA in the late 1980s. Turner recounts previous story generation programs, such as, prominently, Meehan's TALESPIN system, which treated stories as problem solving simulations. Turner's criticism of TALESPIN is that the stories it produced were wholly uninteresting and unmotivated. In writing MINSTREL, Turner wanted to model several layers of storytelling planning tasks, such as the application of story themes, i.e. morals; drama techniques and devices (e.g. suspense, foreshadowing, pacing, characterisation, dialogue, description, tragedy); and the introduction of creativity. One of the canonical stories generated by MINSTREL, called "The Vengeful Princess," illustrates the complexity of the task at hand.
Being a descendant of Schank and Dyer, Turner's view of the modeling world is understandably dominated by the paradigms of scripts, planning, problem solving, and case-based reasoning. In fact, it is Turner's thesis that every cognitive task (including creativity) can be viewed as problem solving. One of the well-known traps of case-based reasoning is that the discretisation of "cases" seems to encourage rote learning and rote application of knowledge. In my opinion, this is rather problematical to the modeling of creativity which my intuition tells me is somehow more fluid, well-connected, and parallel (perhaps an associationist/k-line layer would help?) than CBR is good at. To his credit, Turner recognises the rote learning trap of CBR and tries to address it with some success through the introduction of "mutations."
MINSTREL is set up as a generic case-based problem solver instantiated several times to solve different granularities of the creative authoring problem. From highest to lowest granularity, these subgoals are: theme-selection, consistency-maintainence, dramatic-writing-techniques, and linguistic-presentation. Turner seems quite proud that there is a single generic mechanism that handles all of these subgoals. At each granularity, creativity is interjected in the form of mutations of the problem description. Let me illustrate this idea with an example at the theme-selection granularity. Suppose I have an episodic memory of stories about business situations, and I wanted to adapt one of these stories to the domain of King Arthur. Each story can be represented as a network of frames or a script, with slots like: action, actor, instrument, etc. These features are used to identify the story and provide a way to index stories. (Memory research Tulving has previously argued for the great diversity and specificity of encoded features for retrieving memories in human recall.) In MINSTREL, the case-based reasoner takes a problem description with features from King Arthur, and tries first to look for a rote match in episodic memory. If none exists, usually CBR systems would fail, but Turner's suggestion is to recursively mutate and massage the problem description until a match can be found. For example, change the character "Sir Galahad" to "boss" and "princess" to "employee" or other mutations. What's unclear from the book is the extent of rigour applied to deciding on mutations, but my guess is that it can only be as rigourous as the amount of external knowledge we have about how to make cross-domain analogical mappings (on this point, cf. Fauconnier and (Mark) Turner's work on Conceptual Blending, which specifies metaphorical binding conditions). Turner calls the recursive failure-driven mutation of problem descriptions "creativity." I am skeptical that people are creative in the same way, because at the very least, people are fit at creating sound cross-domain analogical mappings between discourses like business and King Arthur. Turner recognises that some mutations are nonsensical, but instead of proposing more knowledge to assess the soundness of mutations (e.g. Gentner and Forbus' structure-mapping idea), he hand-wavishly excuses these errors by arguing that people too make creativity errors (my position is that while people err, they don't err in such a fundamentally nonsensical way because people are good at things like context and analogy).
One prominent variation on the creativity-as-mutation idea is Turner's suggestion for an imaginative memory. An imaginative memory is created by applying mutations to the individual episodic memories rather than to the problem description. Again, my complaint is that I feel it's unsatisfactory to argue that imagination is simply a mutation operation over episodes without specifying any principles for mutation. A mutation can be a random change, or a complicated context-sensitive inference. My belief is that it is far more complex and I wish there was more dedicated to its specification in Turner's book. My storytelling program MAKEBELIEVE (Liu & Singh, 2002) introduced variation at the lexical level by substitutions using verb alternation classes and semantic synsets -- a rather naive thing to do -- but MINSTREL seems to also use naive mutation for its model of creativity.
As for the rest of the book, there are some fruits to be had. Several interesting thematic goals were implemented, i.e. Planning-Advice-Themes (PATs), and four transformations are randomly applied to each theme to "invent" new themes, namely, generalisation, specialisation, mutation, and recombination. The major role of PAT transformations seems to be to help MINSTREL appear to be less deterministic and hand-crafted. Like creative mutation, there is no way to critique or assess the soundness or goodness of each transformation, because perhaps very broad commonsense knowledge would be required for such a task. MINSTREL implements only one real critic, the boredom assessment, which measures boredom as the ratio of recalled solutions to creative solutions. Perhaps this is the only critic which can be implemented without external knowledge.
The rest of the book is dedicated to discussion of the other story planning tasks such as the introduction of explanatory text to segue from one vignette to another and so on. MINSTREL's evaluation amounts to a quantitative re-enunciation of the various inventories of goals and themes and such, and there is also a limited human evaluation of story quality. My sense is that evaluating these sorts of creative pursuits seems quite difficult, especially with systems like MINSTREL that are so brittle and narrow in their storytelling abilities.
In conclusion, The Creative Process is interesting insofar as it attempts to model two interesting high-level cognitive phenomena of storytelling and creativity. MINSTREL is novel relative to the prior work on story generation because for the first time, there is serious treatment given to author-level goals. I would hesitate to call the proposed model of creativity as anything close to profound because it is implemented as small random variations in a very brittle, narrow, hand-crafted language of scripts and schemas. And really, without broad knowledge of the world to draw from, it is impossible to really emulate the human creative process. I am slightly erked that instead of fessing up to this, Turner back-justifies MINSTREL's creative mutation mechanisms with pseudo-truisms like 'creativity amounts to many small changes' and 'MINSTREL makes creative errors but so do people.' This is reminiscent of the academic context of early AI when researchers thought that humans could be emulated with one or two simple structures and not very much knowledge at all. Only now are we beginning to admit the importance of broad knowledge in all cognitive pursuits.